A pair of data points landed this week that together describe a sharper teen-AI story than the usual cultural-anxiety coverage. Drexel University researchers analyzed more than 300 Reddit posts from users who identified as 13 to 17 years old and had written specifically about dependency on Character.AI; the teens themselves self-reported six classic behavioral-addiction criteria including withdrawal, relapse, and emotional conflict. About a quarter of the posts described using the chatbot for emotional support: loneliness, mental-health struggles, processing distress they were not comfortable sharing with real people. In parallel, a Gallup poll released April 9 showed Gen Z excitement about AI dropped from 36 percent last year to 22 percent in 2026, with 38 percent saying AI does more harm than good for creativity and 42 percent saying the same about critical thinking. Futurism covered both with commentary-driven framing, but the underlying research deserves to be taken on its own terms.
What the Drexel methodology actually shows is narrower and more useful than the headline suggests. It is not a population-level prevalence study; it is a content analysis of users who self-selected to post about dependency, which tells you what the failure mode looks like when it occurs, not how often it occurs. The failure mode itself matters. Character.AI's design surface (unlimited availability, persistent personas, emotional-register matching, conversational nudging toward continued engagement) is precisely the surface that produces the six-criterion pattern the researchers identified. This is not unique to Character.AI. Replika, Xiaoice, Snapchat's My AI, and anyone else shipping companion-framed chat products is optimizing for the same engagement metrics, and the same design choices will produce the same dependency patterns in the subset of users most susceptible to them. The Gallup data is separate: it is population-level sentiment, not specific to any product, and the drop is structural rather than event-driven.
This is approximately the "teens and Instagram, circa 2019" moment for consumer AI. The research literature has started producing specific, methodologically-defensible findings; sentiment data shows a measurable drop in optimism among the target demographic; and the design choices under scrutiny are the same ones that drive the products' commercial traction. The historical pattern from social media is instructive. Regulation lagged evidence by about four years, platforms' own self-regulation lagged evidence by about two, and the companies that aged best were the ones that shipped genuine constraints early (screen-time dashboards, teen-mode defaults, age-gated content), not the ones that issued safety blog posts. Consumer-AI companies in 2026 have a chance to skip most of that lag. Most will not take it, because engagement metrics and retention cohorts do not reward it. The ones who do will be better positioned when the inevitable regulatory and advertiser pressure arrives.
For anyone building consumer-facing AI with significant teen or young-adult usage, the actionable signal is concrete. First, assume the Drexel-style dependency pattern is latent in your product and measure for it: session-length distributions, returning-user cadence, emotional-support conversation prevalence. If your tail users look like the Drexel tail users, you have the problem whether or not you have acknowledged it internally. Second, the design levers that reduce dependency, friction on re-engagement, capped sessions, explicit mental-health redirects, age-gated companion features, also reduce engagement-metric numbers. Making them defaults is a choice the business will push back on; the history of social media suggests the push-back loses eventually and wins on the timeline that matters least. Third, the Gallup trend matters for go-to-market: pitching Gen Z on "AI makes you more creative" now lands in a population where 38 percent believe the opposite. The marketing copy has to change before the product does. For builders targeting older or professional segments, these numbers are less directly relevant, but the downstream regulatory posture will spill over into everyone's terms of service.